Slopsift

Word-list linters tell you ‘delve’ appeared. SlopSift tells you why the sentence is structurally slop. The tool runs a dependency parser — trained from a compact pretrained English encoder using structured distillation and 50 targeted examples — to map every token’s part of speech and grammatical relationship. Rules then inspect that graph for constructions that signal AI-pattern writing: canned outlines repeated across paragraphs, unsupported-certainty claims, vague attribution, and actorless passives. The CLI accepts a glob, lints Markdown and code comments, and emits ESLint-shaped JSON for CI. The browser demo updates on pause, no account required.

The model is the differentiating detail. The vendor states SlopSift is not an API wrapper — they trained a 16 MiB quantized ONNX model specifically for parts-of-speech tagging and dependency parsing, with separate template families held out for evaluation. That model runs on-device in both Node and browser WebAssembly, which means the draft never leaves the machine. For teams in regulated industries or with strict data-handling policies, that is not a nice-to-have; it is the only architecture that clears the review.

SlopSift fits cleanly into a writing-assist loop where a human edits AI-generated drafts and wants a linter that explains what it caught and why — not just a confidence score from a black-box detector. It also fits as a skill inside a coding agent, letting the agent run the real linter and interpret structured findings rather than guessing at style. The wall appears at rule coverage: the documented rule set covers a defined set of structural tells, and teams whose writing problems fall outside that set have no documented path to extend the rule authoring themselves.

Install runs via `bunx slopsift .` with no prior setup. The `–format json` flag produces machine-readable output; `–level info` exposes the full candidate pile including notes. Findings are severity-graded: errors are strong tells, warnings need attention, notes are candidates for writer review — keeping editorial judgment with the human rather than collapsing everything to a pass/fail.